Training-FreeDetector-to-Promptable-Segmenter Integration for Stacked Cartons: Prompt Accuracy–Efficiency Trade-Offs

Detector-assisted promptable segmentation can reuse box annotations, but prompt design creates an accuracy–efficiency trade-off in dense scenes. We study a training-free interface in which a detector trained with bounding-box annotations localizes stacked cartons and a frozen segmenter generates masks. Four interfaces are evaluated: Point-Single, Box, ambiguity-aware Point-Max, and Grid-Iter with sparse positive points and one mask-logit feedback pass. On a leakage-audited split, COCO mask AP, fixed-threshold Hungarian-matched metrics, paired bootstrap uncertainty, and synchronized latency were measured. On 773 test images, Box and Grid-Iter achieved mask AP values of 0.8984 and 0.8988. Their paired AP difference was 0.0007 (95% CI, −0.0035 to 0.0047), while matched-mIoU and F1 intervals also included zero. Box required 604.67 ms/image versus 677.33 ms/image for Grid-Iter. A YOLO26s–SAM 3 Box reference achieved a mask AP value of 0.9115 at 320.74 ms/image, whereas mask-supervised YOLOv9c-seg achieved a value of 0.9235 AP on the same cleaned split. The results indicate a metric-dependent Pareto trade-off rather than overall Grid-Iter superiority and position detector-to-segmenter prompting as a reproducible annotation–accuracy option for densely stacked cartons.